Prima::gp-problems(3) Problems, questionable or intricate topics in 2-D Graphics

Introduction

One of the most important goals of the Prima project is portability between different operating systems. Independently to efforts in keeping Prima internal code that it behaves more or less identically on different platforms, it is always possible to write non-portable and platform-dependent code. Here are some guidelines and suggestions for 2-D graphics programming.

Minimal display capabilities

A compliant display is expected to have minimal set of capabilities, that programmer can rely upon. Following items are guaranteedly supported by Prima:
Minimal capabilities
Distinct black and white colors

Line widths 0 and 1

One monospaced font

Solid fill

rop::Copy and rop::NoOper

Plotting primitives
SetPixel,GetPixel

Line,PolyLine,PolyLines

Ellipse,Arc,Chord,Sector

Rectangle

FillPoly

FillEllipse,FillChord,FillSector

TextOut

PutImage,GetImage

Information services
GetTextWidth,GetFontMetrics,GetCharacterABCWidths

GetImageBitsLayout

Properties
color

backColor

rop

backRop

lineWidth

lineJoin

lineStyle

fillPattern

fillPolyWinding

textOpaque

clipRect

All these properties must be present, however it is not required for them to be changeable. Even if an underlying platform-specific code can only support one mode for a property, it have to follow all obligations for the mode. For example, if platform supports full functionality for black color but limited functionality for the other colors, the wrapping code should not allow color property to be writable then.

Inevident issues

Colors

Black and white colors on paletted displays
Due the fact that paletted displays employ indexed color representation, 'black' and 'white' indices are not always 0 and 2^n-1, so result of raster image operations may look garbled (X). Win32 and OS/2 protect themselves from this condition by forcing white to be the last color in the system palette.

Example: if white color on 8-bit display occupies palette index 15 then desired masking effect wouldn't work for xoring transparent areas with cl::White.

Workaround: Use two special color constants cl::Clear and cl::Set, that represent all zeros and all ones values for bit-sensitive raster operations.

Black might be not 0, and white not 0xffffff
This inevident issue happens mostly on 15- and 16-bits pixel displays. Internal color representation for the white color on a 15-color display ( assuming R,G and B are 5-bits fields) is

 11111000 11111000 11111000
 --R----- --G----- --B-----

that equals to 0xf8f8f8. (All)

Advise: do not check for 'blackness' and 'whiteness' merely by comparing a pixel value.

Pixel value coding
Status: internal

It is not checked how does Prima behave when a pixel value and a platform integer use different bit and/or byte priority (X).

Filled shapes

Dithering
If a non-solid pattern is selected and a background and/or a foreground color cannot be drawn as a solid, the correct rendering requires correspondingly 3 or 4 colors. Some rendering engines (Win9X) fail to produce correct results.
Overfill effect
In complex shapes ( FillPoly, for example) the platform renderer can fill certain areas two or more times. Whereas the effect is not noticeable with rop::CopyPut, the other raster operations (like rop::Xor) produce incorrect picture. (OS/2)

NB - has nothing in common with the fill winding rule.

Workaround: Do not use raster operations with complex filled shapes

Pattern offset
For a widget that contains a pattern-filled shape, its picture will be always garbled after scrolling, because it is impossible to provide an algorithm for a correct rendering without a prior knowledge of the widget nature. (All)

Workaround: Do not use patterned backgrounds. Since the same effect is visible on dithered backgrounds, routine check for pure color might be applied.

Lines

Line caps over patterned styles
It is not clear, whether gaps between dashes should be a multiple to a line width or not. For example, lp::DotDot looks almost as a solid line when lineWidth is over 10 if the first (non-multiple) tactic is chosen. From the other hand it is hardly possible to predict the plotting strategy from a high-level code. The problem is related more to Prima design rather than to a platform-specific code. (All)

Workaround: use predefined patterns (lp::XXX)

Line joins
Joint areas may be drawn two (or more) times - the problem emerges if logical ROP (rop::Xor) is chosen.(OS/2)
Dithering
Dithering might be not used for line plotting. (Win9X)

Fonts

Font metric inconsistency
A font is loaded by request with one size, but claims another afterwards.(OS/2, X).

Impact: system-dependent font description may not match to Prima's.

Advise: do not try to deduce Prima font metrics from system-dependent ones and vice versa.

Transparent plotting
No internal function for drawing transparent bitmaps (like fonts). Therefore, if a font emulation is desired, special ROPs cannot be reproduced. (OS/2, Win9X, WinNT)

Impact: font emulation is laborsome, primarily because the glyphs have to be plotted by consequential anding and xoring a bitmap. Full spectrum of the raster operations cannot be achieved with this approach.

Kerning
Prima do not use text kernings, nor encourages underlying platform-specific code to use it - primarily because of its complexity. From the other hand, sometimes glyph position cannot be determined correctly if no information for the text kerning is provided. (Win9X)
Fractional text position
If the font glyphs have fractional widths, it might be observed that letters may change their position in a string.

Example: A set of glyphs has width of 8.6 pixels for each symbol. If the string ``abcd'' is drawn at position 0, then black part of ``d'' starts at 25th pixel, but if ``cd'' is drawn at 17th, as it supposed to be if the integer arithmetics is used, it starts at 24th pixel. (OS/2)

Solution: Do not rely to Drawable::get_text_width information, because it always returns integer value, but to Drawable::get_font_abc, which returns real values.

Text background
If a text is drawn with non-CopyPut raster operation, text background is not expected to be mixed with symbols - however this is hardly reachable, so results differs for different platforms.

Text background may be only drawn with pure ( non-dithered ) color (Win9X,WinNT) - but this is (arguably) a more correct behavior.

Advise: Do not use ::rop2 and text background for special effects

Internal platform features
Font change notification is not provided. (X, OS/2)

Raster fonts cannot be synthesized (OS/2, partly X)

Raster operations ( ROPs)

Background raster operations are not supported (X,Win9X,WinNT) and foreground ROPs have limited number of modes (OS/2,X). Not all ROPs can be emulated for certain primitives, like fonts, complex shapes, and patterned shapes.

It is yet unclear which primitives have to support ROPs, - like FloodFill and SetPixel. Behavior of the current implementation is that they do not.

Arcs

Platforms tend to produce different results for angles outside 0 and 2pi. Although Prima assures that correct plotting would be performed for any angle, minor inconsistencies may be noticed. If emulating, note that 2 and 4-pi arcs are not the same - for example, they look differently with rop::Xor.

Palettes

Static palettes
Some displays are unable to change their hardware palette, so detecting 8- or 4- bits display doesn't automatically mean that palette is writable.(X)
Widget::palette
Widget::palette property is used for explicit declaration of extra color needs for a widget. The request might be satisfacted in different ways, or might not at all. It is advisable not to rely onto platform behavior for the palette operations.
Dynamic palette change
It is possible (usually on 8-bits displays) for a display to change asynchronously its hardware palette in order to process different color requests. All platforms behave differently.

Win9X/WinNT - only one top-level window at a time and its direct children ( not ::clipOwner(0)) can benefit from using Widget::palette. System palette is switched every time as different windows moved to the front.

OS/2 - not implemented, but in principle the same as under win32.

X - Any application can easily ruin system color table. Since this behavior is such by design, no workaround can be applied here.

Bitmaps

Invalid scaling
Scaling is invalid (Win9X) or not supported (X). Common mistake is to not take into an account the fractional pixels that appear when the scaling factor is more than 1. This mistake can be observed in Win9X.

Workaround: none

Large scale factors
Request for drawing a bitmap might fail if large scaling factor is selected. (OS/2,Win9X,WinNT). This effect is obviously due that fact that these platforms scale the bitmap into a memory before the plotting takes place.

Platform-specific peculiarities

OS/2

Some ROPs are ambiguous - SRCTRANSPARENT, for example. Some times they work, some times they don't. The particular behavior depends on a video driver.

Circles cannot be drawn using an even diameter.

Fast GDI operations on HWND_DESKTOP may be delayed, thus GetPixel may return invalid pixel values.

Windows 9X

Amount of GDI objects can not exceed some unknown threshold - experiments show that 128 objects is safe enough.

No transformations.

Color cursor creation routine is broken.

Filled shapes are broken.

X

No transformations

No bitmap scaling

No font rotation

No GetPixel, FloodFill ( along with some other primitives)

White is not 2^n-1 on n-bit paletted displays (tested on XFree86).

Filled shapes are broken.

Color bitmaps cannot be drawn onto mono bitmaps.

Implementation notes

OS/2

Palettes are not implemented

Win32

Plotting speed of DeviceBitmaps is somewhat less on 8-bit displays than Images and Icons. It is because DeviceBitmaps are bound to their original palette, so putting a DeviceBitmap onto different palette drawable employs inefficient algorithms in order to provide correct results.

X

Image that was first drawn on a paletted Drawable always seen in 8 colors if drawn afterwards on a Drawable with the different palette. That is because the image has special cache in display pixel format, but cache refresh on every PutImage call is absolutely inappropriate (although technically possible). It is planned to fix the problem by checking the palette difference for every PutImage invocation. NB - the effect is seen on dynamic color displays only.

AUTHOR

Dmitry Karasik, <[email protected]>.